Bak-Coleman, J., O’Connor, C., Bergstrom, C., & West, J. (2025). The risks of industry influence in tech research. arXiv [cs.SI].
Summary
This Perspective argues that the technology industry exerts substantial, underappreciated influence over the scientific study of its own products, following patterns established historically by the tobacco, fossil fuel, asbestos, and soda industries — but amplified by tech’s unique control over the data, funding, and access required to study platforms. Drawing on the agnotology and political economy of science literatures, the authors contend that existing scientific safeguards (IRB review, conflict-of-interest disclosure, open science practices) are inadequate and can paradoxically disadvantage independent researchers relative to industry-affiliated ones. They catalog the mechanisms of industry influence and propose reforms spanning journals, funders, individual scientists, and governments.
Key Contributions
- Provides a unified taxonomy of mechanisms of tech-industry influence: burying internal research, selective publishing, design bias, selective/industrial funding, inhibition of independent research, and performative collaboration.
- Extends established critiques of industry-science conflicts (Oreskes & Conway, Holman & Bruner) into the domain of digital information technologies and computational social science.
- Identifies how Open Science reforms (preregistration, large samples, causal RCTs) can be co-opted or asymmetrically applied to favor industry-produced work.
- Offers concrete policy proposals: mandatory IRB review for industry research, stronger COI auditing, mandated data access (building on the EU DSA), required disclosure of internal studies, dedicated independent funding, and an IPCC-like Intergovernmental Panel on Information Technology.
- Provides practical guidance for academics weighing industry collaborations, focused on design bias, “first-peek” risks, and data integrity.
Methods
A conceptual and argumentative Perspective synthesizing prior literature on industry-science conflicts. The analysis is case-based, drawing on historical examples (tobacco, asbestos, fossil fuels, soda) alongside contemporary tech examples (Meta, X/Twitter, Google, Coca-Cola collaborations). It examines leaked internal documents (the Frances Haugen and 2025 Meta whistleblower caches), compares statistical and methodological choices across Meta-affiliated publications, and includes an informal OpenAlex bibliometric check of Meta-funded researchers versus disclosed Meta funding.
Findings
- Meta and other companies have suppressed or deleted internal research on harms to children, teen mental health, and VR safety.
- Industry-affiliated studies tend to apply lenient statistical criteria (e.g., p<.1) when measuring benefits but stricter corrections when measuring harms, yielding asymmetric conclusions from comparable effect sizes.
- RCTs of short platform breaks systematically underestimate long-term, cumulative, and societal-scale harms.
- Independent data access has eroded sharply (CrowdTangle shutdown, X API restrictions, legal threats against NYU Ad Observatory and CCDH).
- The Meta/Social Science One collaboration exhibited design bias, undisclosed platform interventions during study periods, missing user data, and deviations from preregistration.
- Many tech-funded academics fail to disclose industry ties: ~1000+ Meta-funded researchers but only 472 OpenAlex publications acknowledging Meta funding.
- Industry research is frequently exempted from IRB review by journals, an asymmetry that hinders independent replication.
Connections
This paper anchors debates over platform data access and independent research infrastructure; its critique of eroding data access and the limits of the EU DSA connects directly to work assessing data-access regimes and research APIs such as Rieder2025-ju, Rieder2026-pp, and Davo2026-… The authors’ concern with the reliability and independence of the platform-study evidence base relates to their own companion work in Bak-Coleman2026-mk, as well as broader methodological and data-quality critiques in Efstratiou2025-gs and Murtfeldt2025-wu.